Single cell signal path reconstruction method, system, equipment and medium

Abstract

The invention relates to a single cell signal road reconstruction method, system and device and a medium. The method comprises the following steps: extracting a sub-network from a global signal reference network based on an intersection of a highly expressed gene and a hypervariable gene of each cell, and generating a cell specific signal network; inputting the cell specificity signal network into a Transform model to generate a consensus signal network; defining the consensus signal network corresponding to the source cell as a source cell consensus network, and defining the consensus signal network corresponding to the target cell as a target cell consensus network; and on the basis of a ligand-receptor knowledge base, searching a path between a ligand in the source cell consensus network and a receptor in the target cell consensus network, and generating a cross-cell signal path list. By adopting the method, the intercellular communication mechanism can be accurately captured, and the systematic understanding of intercellular signals is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of deep learning, and in particular relates to a method, system, device and medium for reconstructing single-cell signaling pathways. Background Technology

[0002] With the rapid development of single-cell RNA sequencing technology, scientists have been able to observe complex tissue microenvironments at single-cell resolution, bringing unprecedented opportunities for elucidating intercellular communication. Against this backdrop, analytical methods based on ligand-receptor co-expression have emerged. These methods infer potential intercellular interactions by statistically analyzing the expression patterns of ligand and receptor genes in adjacent cell populations. Such methods typically rely on a predefined database of ligand-receptor pairs. The analytical process first calculates the average expression levels of ligands and receptors across different cell types, then uses statistical tests to assess the significance of ligand-receptor pair expression correlations, thereby screening for cell type pairs that may be communicating and the ligand-receptor signaling units they use. However, this current mainstream approach has a fundamental limitation: it treats intercellular communication as an isolated event. The analytical results can only indicate that cell A may have sent a signal to a receptor in cell B via a certain ligand, but cannot answer the core biological question of what downstream signal transduction cascades were triggered within cell B after the signal was received, and which target genes were ultimately regulated, thus affecting the functional state of cell B. This analytical approach cannot fully reconstruct the pathways of intercellular communication, greatly limiting the systematic and causal understanding of intercellular communication mechanisms. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, system, device, and medium for single-cell signaling pathway reconstruction that can completely reconstruct the interpretable pathway from source cell to target cell, addressing the aforementioned technical problems.

[0004] In a first aspect, this application provides a method for reconstructing a single-cell signaling pathway, comprising:

[0005] Based on the intersection of highly expressed genes and hypervariable genes in each cell, a sub-network is extracted from the global signal reference network to generate a cell-specific signal network; the nodes of the cell-specific signal network are hypervariable genes, and the edges are the interactions between hypervariable genes.

[0006] The cell-specific signaling network is input into the graph Transformer model to generate a consensus signaling network;

[0007] The consensus signal network of the corresponding source cell is defined as the source cell consensus network, and the consensus signal network of the corresponding target cell is defined as the target cell consensus network.

[0008] Based on the ligand-receptor knowledge base, the pathways between ligands in the source cell consensus network and receptors in the target cell consensus network are searched to generate a list of cross-cell signaling pathways.

[0009] Furthermore, the graph Transformer model generates the consensus signaling network through the following method:

[0010] Each node in the cell-specific signaling network is downsampled into a low-dimensional value to obtain the encoded node embedding vector; and the encoded node embedding vectors are aggregated into a cell embedding vector.

[0011] Based on cell type annotation information, cell embedding vectors belonging to the same cell type are aggregated to obtain consensus cell embedding vectors.

[0012] Based on the consensus cell embedding vector, a low-dimensional vector is generated for each node to obtain the inference node embedding vector; the inference node embedding vector represents the functional context of the node in the typical state of cell type.

[0013] Based on the embedding vectors of the inference nodes, the connection probabilities between each node are calculated using the following formula to obtain the probabilistic adjacency matrix:

[0014]

[0015] in, Let be the probability that there is an edge between node i and node j. Let i be the node embedding vector. Let j be the node embedding vector;

[0016] The consensus signal network is obtained by eliminating the interactions between nodes whose connection probability is less than a preset threshold.

[0017] Furthermore, the graph Transformer model is trained using the following method:

[0018] The training set is input into the encoder of the graph Transformer model to generate encoding node embedding vectors and cell embedding vectors; the training set includes cell-specific signaling networks and corresponding cell type labels.

[0019] The encoded nodes are embedded into vectors and input into the decoder. The connection probability between each node is calculated, and a reconstruction probability matrix is ​​generated.

[0020] The cell embedding vector is input into the classifier to obtain the cell type probability distribution;

[0021] Based on the reconstruction probability matrix, cell type probability distribution, adjacency matrix of cell-specific signal network, and cell type label, multi-task loss calculation is performed to obtain the joint loss value;

[0022] Based on the joint loss value, the parameters of each module of the graph Transformer model are updated to obtain the updated graph Transformer model.

[0023] Furthermore, based on the reconstruction probability matrix, cell type probability distribution, adjacency matrix of cell-specific signaling network, and cell type labels, multi-task loss calculation is performed to obtain the joint loss value, including:

[0024] Based on the reconstruction probability matrix and the adjacency matrix of the cell-specific signaling network, the reconstruction loss is calculated using the following formula:

[0025]

[0026]

[0027] in, For positive sample reconstruction loss, It is the set of all pairs of nodes in the adjacency matrix that have an edge. For negative sample reconstruction loss, It is the set of all pairs of nodes in the adjacency matrix that do not have an edge.

[0028] Based on the cell type probability distribution and cell type labels, the classification loss is calculated using the following formula:

[0029]

[0030] in, The classification loss is represented by C, where C is the total number of cell types. For actual cell type labels, Let c be the probability that the cell belongs to type c.

[0031] The joint loss value is obtained by weighted summation of reconstruction loss and classification loss.

[0032] Furthermore, based on the ligand-receptor knowledge base, pathways between ligands in the source cell consensus network and receptors in the target cell consensus network are searched to generate a list of cross-cell signaling pathways, including:

[0033] The ligand-receptor knowledge base is queried to match all ligand-receptor pairs that meet the preset conditions, resulting in a list of candidate ligand-receptor pairs. The preset conditions are that the ligand gene exists in a node of the source cell consensus network and the receptor gene exists in a node of the target cell consensus network.

[0034] For each receptor in the ligand-receptor pair list, a path leading to the transcription factor node is searched in the target cell consensus network to obtain a receptor-transcription factor path list.

[0035] In the target cell consensus network, all target genes regulated by transcription factors are queried and extracted to obtain the transcription factor-target gene pathway; and the transcription factor-target gene pathway is added to the receptor-transcription factor pathway list to obtain an expanded signaling pathway list.

[0036] By concatenating the candidate ligand-receptor pair list and the signaling pathway list, a list of transcellular signaling pathways is obtained.

[0037] Furthermore, for each receptor in the ligand-receptor pair list, a path to the transcription factor node is searched in the target cell consensus network to obtain a receptor-transcription factor path list, including:

[0038] Extract all receptor genes located on the target cell side from the candidate ligand-receptor pair list to generate a receptor gene list;

[0039] Starting with each receptor gene in the receptor gene list as a starting node, the target cell consensus network is traversed, and the direct neighboring nodes of the initial node are recorded to obtain a set of potential paths.

[0040] By filtering out the potential pathways whose last node is a transcription factor, a list of receptor-transcription factor pathways is obtained.

[0041] Furthermore, the global signal reference network is constructed using the following method:

[0042] Obtain raw data on cell signaling interactions; and extract the participants and corresponding relationship types from the raw data to obtain a set of interaction data; the raw data includes at least one of the following: transcription factor-target gene regulatory relationships, protein-protein interactions, metabolic responses, signal transduction events, and ligand-receptor pairing information;

[0043] The action data set is standardized to obtain the standard action set;

[0044] Based on the preset reliability criteria, records that do not meet the reliability requirements in the standard action set are removed to obtain the reliable action set;

[0045] Based on a reliable set of interactions, each gene or protein node and each interaction are treated as an edge between nodes, and the edges are labeled according to the type of interaction to obtain a preliminary signal reference network.

[0046] The initial signal reference network is divided into intracellular signal modules and intercellular signal modules to obtain a global signal reference network; the intracellular signal modules are used to characterize intracellular signal cascades; and the intercellular signal modules are used to characterize ligand-receptor interaction pairs.

[0047] Secondly, this application also provides a single-cell signaling pathway reconstruction system, comprising:

[0048] The generation module is used to extract subnetworks from the global signal reference network based on the intersection of highly expressed genes and hypervariable genes in each cell, and generate cell-specific signal networks. The nodes of the cell-specific signal network are hypervariable genes, and the edges are the interactions between hypervariable genes.

[0049] The consensus module is used to input cell-specific signaling networks into the graph Transformer model and generate a consensus signaling network.

[0050] The definition module is used to define the consensus signal network of the corresponding source cell as the source cell consensus network and the consensus signal network of the corresponding target cell as the target cell consensus network.

[0051] The path module is used to search for pathways between ligands in the source cell consensus network and receptors in the target cell consensus network based on the ligand-receptor knowledge base, and generate a list of cross-cell signaling pathways.

[0052] Thirdly, this application also provides a computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any step of the method provided in the first aspect of this application.

[0053] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any step of the method provided in the first aspect of this application.

[0054] The aforementioned method, system, device, and medium for reconstructing single-cell signaling pathways extracts sub-networks from a global signaling reference network based on the intersection of highly expressed and hypervariable genes in each cell, generating a cell-specific signaling network. Nodes in the cell-specific signaling network represent hypervariable genes, and edges represent interactions between hypervariable genes. The cell-specific signaling network is input into a graph Transformer model to generate a consensus signaling network. The consensus signaling network corresponding to the source cell is defined as the source cell consensus network, and the consensus signaling network corresponding to the target cell is defined as the target cell consensus network. Based on a ligand-receptor knowledge base, pathways between ligands in the source cell consensus network and receptors in the target cell consensus network are searched to generate a list of cross-cell signaling pathways. This method can completely construct an interpretable pathway from ligand secretion in the source cell to receptor reception, internal signal transduction, transcription factor activation, and ultimately, the regulation of specific target genes in the target cell, greatly enhancing the depth and interpretability of intercellular communication research. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram of the process of a single-cell signaling pathway reconstruction method provided in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the structure of a single-cell signaling pathway reconstruction system provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1 As shown, a method for reconstructing a single-cell signaling pathway is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0060] Step 101: Based on the intersection of highly expressed genes and hypervariable genes in each cell, extract sub-networks from the global signal reference network to generate a cell-specific signal network; the nodes of the cell-specific signal network are hypervariable genes, and the edges are the interactions between hypervariable genes.

[0061] Specifically, the cell is the basic unit of analysis in scRNA-seq (Single-cell RNA Sequencing) data, and each cell possesses its unique gene expression profile. Highly expressed genes are those genes with the highest expression levels within a specific cell, representing the most active genes in that cell's current state. Hypervariable genes are those genes with the largest expression variance across the entire cell population, exhibiting significant expression differences between different cells and rich in information related to cell state and functional heterogeneity; they are the core of single-cell analysis. The global signaling reference network is a pre-constructed, comprehensive, large-scale knowledge base integrating information from multiple databases, encompassing a wide range of known biomolecular interactions, such as protein-protein interactions, transcription factor-regulated target genes, signal transduction pathways, and ligand-receptor relationships. As a background knowledge base, it forms the basis for all analyses. Subnetworks are portions of the network extracted from the global signaling reference network based on the gene list. The cell-specific signaling network is a graph structure, where nodes are the intersection genes belonging to both the cell's highly expressed genes and the global hypervariable genes; edges represent the interaction relationships between these node genes within the global signaling reference network. For example, for the cell being processed, the terminal calculates the top 500 genes by expression level and retrieves the top 5000 genes by variance in the whole genome from the entire dataset. The intersection of these two gene lists is then identified. The significance of this intersection gene set lies in the fact that these genes are both highly active in the current cell and exhibit significant differences between different cells. Therefore, they are likely to be key signaling molecules that determine the functional state of the cell. Using this intersection gene list as a screening criterion, the terminal queries and matches the global signal reference network, retaining all interactions between the two endpoints of the reference network that belong to this intersection gene set. These selected genes and interactions are then used to construct an independent, smaller network, which is the cell-specific signaling network for this cell.

[0062] Step 102: Input the cell-specific signaling network into the graph Transformer model to generate a consensus signaling network.

[0063] Among them, the Graph Transformer model is a deep learning model specifically designed for processing graph-structured data. It can learn the numerical representation of each node in the graph, while considering the node's own characteristics and its connections with other nodes based on an attention mechanism. The consensus signal network is a typical signal network representing the entire cell type, with individual cell-specific noise removed. Its nodes are the same as the input network, but the presence and strength of edges are inferred based on the common patterns of all cells within that cell type. The terminal inputs cell-specific signal networks one by one into a pre-trained graph Transformer model. The model processes each input network, analyzes its structure, and generates a low-dimensional, numerical embedding vector for each gene node. This vector encodes the functional context information of the gene in a specific cell network. The model aggregates node information from all cells within the same cell type. For each gene node, it calculates the aggregated value of its corresponding embedding vector in all cells of that cell type, generating a consensus node embedding vector. This consensus vector represents the functional role of the gene in a typical state of that cell type. Based on these consensus node embedding vectors, the model calculates the probability of a connection between any two gene nodes by comparing the similarity of their embedding vectors. A probability threshold is set, and only edges with a connection probability higher than the threshold are retained, while low-probability edges are discarded, thus generating a consensus signal network.

[0064] Step 103: Define the consensus signal network of the corresponding source cell as the source cell consensus network, and define the consensus signal network of the corresponding target cell as the target cell consensus network.

[0065] In this context, the source cell is the cell type that sends the signal in the inter-cell communication to be analyzed. The target cell is the cell type that receives the signal in the same inter-cell communication. For example, if an immune cell sends an activation signal to a cancer cell, the immune cell is the source cell, and the cancer cell is the target cell. The source cell consensus network is a consensus signal network corresponding to the source cell type. The target cell consensus network is a consensus signal network corresponding to the target cell type. Based on the biological problem or analytical needs, the user specifies which is the source cell type and which is the target cell type. The terminal then marks the generated consensus signal network corresponding to the source cell type as the source cell consensus network and the consensus signal network corresponding to the target cell type as the target cell consensus network.

[0066] Step 104: Based on the ligand-receptor knowledge base, search for pathways between ligands in the source cell consensus network and receptors in the target cell consensus network to generate a list of transcellular signaling pathways.

[0067] Specifically, a ligand-receptor knowledge base is a pre-organized database containing known ligand-receptor pairs. A ligand is a signaling molecule that can be secreted by a cell or expressed through its membrane, while a receptor is another protein on the cell surface or inside that can specifically recognize the ligand and trigger a response. In a network, a pathway refers to a series of interconnected edges and nodes, specifically a series of continuous interactions starting from the receptor gene, passing through a series of intermediate signaling molecules, and finally reaching the transcription factor. A transcellular signaling pathway list is a list where each record is a complete, hypothetical communication pathway describing the entire process from the ligand in the source cell to the receptor in the target cell, then to signal transduction within the target cell, ultimately leading to changes in the expression of a specific gene. The terminal queries the ligand-receptor knowledge base to filter out all ligand-receptor pairs that meet the criteria. The ligand gene must exist in a node of the source cell consensus network, and the receptor gene must exist in a node of the target cell consensus network. For each receptor found, in the target cell consensus network, starting from it, a graph traversal algorithm is used to find all paths that can reach the transcription factor node, limiting the path length to ensure the biological rationality and interpretability of the pathway. In the target cell consensus network, all target genes regulated by the endpoints of the above paths are found. These target genes are taken as the final links in the pathway, connecting the upstream ligand, the matched receptor, the signaling pathway from the receptor to the transcription factor, and the regulatory relationship from the transcription factor to the target gene to form a complete pathway. All the found complete pathways are summarized to generate a list of transcellular signaling pathways.

[0068] This embodiment provides a single-cell signaling pathway reconstruction method. Based on the intersection of highly expressed genes and hypervariable genes in each cell, a sub-network is extracted from the global signaling reference network to generate a cell-specific signaling network. Nodes in the cell-specific signaling network represent hypervariable genes, and edges represent interactions between hypervariable genes. The cell-specific signaling network is input into a graph Transformer model to generate a consensus signaling network. The consensus signaling network corresponding to the source cell is defined as the source cell consensus network, and the consensus signaling network corresponding to the target cell is defined as the target cell consensus network. Based on a ligand-receptor knowledge base, pathways between ligands in the source cell consensus network and receptors in the target cell consensus network are searched to generate a list of cross-cell signaling pathways. Through these methods, a complete and interpretable pathway can be constructed, from ligand secretion in the source cell to receptor reception, internal signal transduction, transcription factor activation, and ultimately the regulation of specific target genes in the target cell. This explains the communication mechanisms under typical cellular conditions, greatly enhancing the depth and interpretability of intercellular communication research.

[0069] In one embodiment, the graph Transformer model generates the consensus signaling network through the following method:

[0070] Step 201: Downsample each node in the cell-specific signaling network to a low-dimensional value to obtain the encoded node embedding vector; and aggregate the encoded node embedding vector into a cell embedding vector.

[0071] Downsampling refers to feature extraction and dimensionality reduction using neural networks, aiming to compress complex node information into a low-dimensional space that retains key features. The encoded node embedding vector is a fixed-length low-dimensional numerical vector representing the functional state of a single gene in a specific cellular environment. It encodes the position and context of the gene within the cell's specific network, i.e., which other genes it interacts with. The cell embedding vector is a fixed-length low-dimensional numerical vector aggregated from the embedding vectors of all nodes in a cell. It represents the overall characteristics of the entire cellular signaling network and is a highly generalized and abstract representation of the cell's functional state. The model processes each node in the cell-specific signaling network. Through multi-layer computation, including self-attention mechanisms, it comprehensively considers its own attributes and information from all its neighboring nodes to generate a low-dimensional, numerical representation for that node—the encoded node embedding vector. This transforms complex network relationships into a computable mathematical form. Once all nodes in a cell have their own embedding vectors, the model uses an aggregation function to merge these vectors into a single cell embedding vector, capturing the most salient and representative patterns in the cellular signaling network.

[0072] Step 202: Based on cell type annotation information, aggregate cell embedding vectors belonging to the same cell type to obtain consensus cell embedding vectors.

[0073] Specifically, cell type annotation information consists of the label of the cell type to which each cell belongs, and this information is known in advance. The consensus cell embedding vector is a single vector obtained by aggregating all cell embedding vectors of the same cell type. It represents the typical functional state of that cell type, reducing the influence of outliers or noise from individual cells. Based on the cell type annotation information, the model selects all cells belonging to the same type, aggregates these vectors using a pruned mean, first removing a certain proportion of maximum and minimum values, and then averaging the remaining values. This effectively resists the influence of extreme outliers and yields a more reliable consensus representation.

[0074] Step 203: Based on the consensus cell embedding vector, generate a low-dimensional vector for each node to obtain the inference node embedding vector; the inference node embedding vector represents the functional context of the node in the typical state of the cell type.

[0075] Specifically, the inference node embedding vector is a low-dimensional numerical vector derived from the consensus cell embedding vector. It represents the functional context of a gene in a typical, average state within that cell type, removing noise specific to individual cells. The model decoder uses the consensus cell embedding vector as global context information. For each gene node, it utilizes this consensus information to generate a completely new embedding vector, namely the inference node embedding vector.

[0076] Step 204: Based on the inference node embedding vectors, calculate the connection probabilities between each node using the following formula to obtain the probabilistic adjacency matrix:

[0077]

[0078] in, Let be the probability that there is an edge between node i and node j. Let i be the node embedding vector. Let be the node embedding vector of node j.

[0079] Specifically, the probabilistic adjacency matrix is ​​a square matrix whose elements are... This is a probability value between 0 and 1, representing the probability that there is an edge between node i and node j in the consensus network. It is used to describe the probabilistic connection structure of the entire network. For any pair of nodes i and j in the model, the embedding vectors of the inference nodes are extracted, and the inner product of these two vectors is calculated. The result of the inner product is a scalar. The larger the value, the more similar and related the directions of the two vectors are in the latent space. The inner product is input into the activation function, which converts the inner product value into a probability value. The higher the inner product value, the closer the converted probability is to 1; the lower the inner product value, the closer the probability is to 0. This calculation is performed on all possible node pairs, filling in a complete probabilistic adjacency matrix.

[0080] Step 205: Eliminate interactions between nodes whose connection probability is less than a preset threshold to obtain the consensus signal network.

[0081] The preset threshold is a manually set probability threshold used to determine whether an edge is reliable enough to be retained in the final network. The consensus signal network is a deterministic graph structure where nodes are genes and edges are interactions between genes, representing the computational inference results of the most core and stable signaling pathways for that cell type. The model traverses each element in the probabilistic adjacency matrix, comparing each probability value with the preset threshold. Edges with probabilities less than the threshold are discarded, as they are considered unreliable and may be noise or atypical interactions. Edges with probabilities greater than or equal to the threshold are retained, as these are considered stable, functionally relevant core interactions in that cell type. All retained edges are output along with the nodes to obtain the consensus signal network.

[0082] This embodiment balances the confidence and sparsity of the consensus signal network through threshold control. By converting the similarity between nodes into connection probabilities, the connection structure of the network is reconstructed from the embedding vector. Aggregation uses pruned mean to effectively resist the influence of extreme outliers, thereby improving the accuracy of information in the consensus signal network.

[0083] In one embodiment, the graph Transformer model is trained using the following method:

[0084] Step 301: Input the training set into the encoder of the graph Transformer model to generate the encoding node embedding vector and cell embedding vector; the training set includes the cell-specific signal network and the corresponding cell type label.

[0085] Specifically, the training set is the dataset used to train the model. Each training sample contains a cell-specific signaling network and an identifier for the cell type. The encoder, a crucial component of the graph Transformer model, reads and understands the input graph data. Composed of multiple graph attention layers and other structures, it captures the complex dependencies between nodes in the graph. The encoded node embedding vector is a numerical representation generated by the encoder for each gene node in the input network. It condenses the node's structural information and functional context within the network, resulting in a low-dimensional, dense vector. The cell embedding vector is a single vector generated by aggregation of the encoded node embedding vectors of all nodes in the cell. It represents the global features of the entire cell signaling network and provides a general description of the cell's functional state. The terminal retrieves a batch of training samples from the training set. Each sample contains a cell-specific signaling network and its cell type label. The network data is input into the model's encoder. The encoder processes the network data layer by layer, calculating attention weights and other information between nodes to generate a novel, semantically rich encoded node embedding vector for each node in the network. The model then uses an aggregation function to combine the embedding vectors of all nodes into a single cell embedding vector.

[0086] Step 302: Embed the encoded nodes into the vector input decoder, calculate the connection probability between each node, and generate the reconstruction probability matrix.

[0087] The decoder is a component of the model whose task is to predict whether there is a connection between nodes based on their embedding representations. For example, it could be an inner product decoder. The reconstruction probability matrix is ​​a matrix where each element is a probability value between 0 and 1, representing the likelihood that an edge exists between node i and node j as predicted by the model. The decoder calculates a connection score for each possible pair of nodes (i,j) in the network. It computes the inner product of the two node vectors, maps this score to the (0,1) interval using an activation function, and interprets it as a probability. After calculating this for all node pairs, the complete reconstruction probability matrix is ​​obtained.

[0088] Step 303: Input the cell embedding vector into the classifier to obtain the cell type probability distribution.

[0089] The classifier is a feedforward neural network that maps the input cell embedding vector to different cell type categories. The cell type probability distribution is a vector with a length equal to the total number of cell types. Each element in the vector represents the probability that the model predicts the cell belongs to a specific cell type, and the sum of the probabilities of all elements is 1. The terminal inputs the cell embedding vector into the classifier, which typically consists of several fully connected layers and activation functions, and outputs the cell type probability distribution.

[0090] Step 304: Based on the reconstruction probability matrix, cell type probability distribution, adjacency matrix of cell-specific signal network, and cell type label, perform multi-task loss calculation to obtain the joint loss value.

[0091] Specifically, the adjacency matrix of the cell-specific signal network is the true connectivity matrix corresponding to the network in the input training set. It is a binary matrix used to indicate which nodes have edges and which do not, serving as the true standard for evaluating the quality of network reconstruction. The cell type label is the true type of the cell and serves as the true standard for evaluating the quality of classification prediction. The joint loss value is a single scalar value, calculated through multi-task loss, reflecting the total error of the model in both reconstruction and classification tasks on the current data. The terminal takes these four elements as input and calculates the input using loss calculation rules that simultaneously consider both network reconstruction and cell classification objectives, obtaining a single numerical result, the joint loss value.

[0092] Step 305: Based on the joint loss value, update the parameters of each module of the graph Transformer model to obtain the updated graph Transformer model.

[0093] Specifically, parameters refer to all adjustable weights and biases in the model, distributed across the encoder, decoder, and classifier. The updated graph Transformer model refers to the new version of the model whose internal state has been optimized after this parameter adjustment. The terminal uses the backpropagation algorithm to calculate the gradient of the loss with respect to each parameter in the model, starting from the joint loss value. The gradient indicates the direction and magnitude of adjustment for each parameter to reduce the loss. The optimizer uses the calculated gradients to actually update all parameters, and this process is repeated on all data in the training set until the joint loss value converges to a low level, indicating that the model has been sufficiently trained.

[0094] This embodiment forces the model to learn a general and powerful feature representation that simultaneously satisfies network reconstruction and cell classification through a multi-task learning framework. This results in a stronger generalization ability than a model that performs either task alone, thereby improving the accuracy of cell signaling pathway reconstruction.

[0095] In one embodiment, multi-task loss calculation is performed based on the reconstruction probability matrix, cell type probability distribution, adjacency matrix of cell-specific signaling network, and cell type labels to obtain a joint loss value, including:

[0096] Step 401: Based on the reconstruction probability matrix and the adjacency matrix of the cell-specific signaling network, calculate the reconstruction loss using the following formula:

[0097]

[0098]

[0099] in, For positive sample reconstruction loss, It is the set of all pairs of nodes in the adjacency matrix that have an edge. For negative sample reconstruction loss, It is the set of all node pairs in the adjacency matrix that do not have an edge.

[0100] Specifically, the reconstruction probability matrix is ​​a matrix where each element represents the probability that a connection exists between node i and node j, as predicted by the model. This matrix represents the model's inference of the network structure. The adjacency matrix of the cell-specific signaling network is a mathematical representation of the input real network; it is a binary matrix where an element of 1 indicates the existence of an edge between the corresponding two nodes, and a value of 0 indicates the absence of an edge. The positive sample set refers to the set of node pairs corresponding to all elements in the adjacency matrix with a value of 1, i.e., all node pairs that are actually connected in the real network. The negative sample set refers to the set of node pairs corresponding to all elements in the adjacency matrix with a value of 0, i.e., all node pairs that are actually not connected in the real network. The positive sample reconstruction loss is a scalar value specifically measuring the model's error in predicting existing edges. It is calculated as the logarithmic mean of the model's predicted probabilities for all truly existing edges, and is negative. The negative sample reconstruction loss is a scalar value specifically measuring the model's error in predicting non-existent edges. It is calculated as the logarithmic mean of the model's predicted probabilities of no edge for all non-existent edges, and is negative. The terminal iterates through every node pair in the positive sample set. For each node pair, it extracts the model-predicted connection probability from the reconstructed probability matrix and calculates the natural logarithm of the probability. When the predicted probability is close to 1, its logarithm is close to 0; when the predicted probability is small, its logarithm will be a large negative number, which, after negation, becomes a large positive number. This process is repeated for all positive samples. Summing the results and dividing by the total number of positive samples yields the average. Taking the negative value gives the reconstruction loss for the positive samples. Iterating through each node pair in the negative sample set, for each pair, calculates the probability of the model predicting no boundary. The natural logarithm of this probability is then calculated. The loss is small when the model is certain of no boundary and makes a perfect prediction; the loss is large when the probability of no boundary is too small. This process is repeated for all negative samples. Summing the results and dividing by the total number of negative samples yields the average value. Taking the negative value gives the negative sample reconstruction loss.

[0101] Step 402: Based on the cell type probability distribution and cell type labels, calculate the classification loss using the following formula:

[0102]

[0103] in, The classification loss is represented by C, where C is the total number of cell types. For actual cell type labels, Let be the probability that the cell belongs to type c.

[0104] Specifically, the cell type probability distribution is a probability vector representing the likelihood that the model considers the current cell to belong to each cell type. The true cell type label refers to the actual cell type identifier provided in the training samples, represented in one-hot encoding form. The classification loss is a scalar value that measures the difference between the cell type probability distribution predicted by the model and the true cell type label. This embodiment uses the standard cross-entropy loss used in classification tasks. The terminal iterates through all possible cell types. For each type, it extracts the true label value and the model-predicted probability value, substitutes them into the formula, and calculates the classification loss.

[0105] Step 403: Weighted summation of reconstruction loss and classification loss to obtain the joint loss value.

[0106] The reconstruction loss is a collective term for the reconstruction losses of positive and negative samples, representing the total error of the model on the unsupervised learning task. The classification loss represents the error of the model on the supervised learning task. The joint loss is a single scalar value, obtained by weighted summation of the reconstruction and classification losses, and serves as a global objective function guiding the update of the model parameters. The terminal assigns weight coefficients to each loss and sums the weighted results to obtain the joint loss value.

[0107] This embodiment controls the learning focus of the model by using different weighted losses, while improving the model's ability to reconstruct networks and classify cells. It guides the cell embedding vectors learned by the model to contain enough information to accurately distinguish different cell types, ensuring that the node embeddings learned by the model can most accurately retain the connection information of the original network, thereby improving the stability of signal pathway reconstruction.

[0108] In one embodiment, based on a ligand-receptor knowledge base, pathways between ligands in the source cell consensus network and receptors in the target cell consensus network are searched to generate a list of cross-cell signaling pathways, including:

[0109] Step 501: Query the ligand-receptor knowledge base, match all ligand-receptor pairs that meet the preset conditions, and obtain a list of candidate ligand-receptor pairs; the preset conditions are that the ligand gene exists in the node of the source cell consensus network and the receptor gene exists in the node of the target cell consensus network.

[0110] The ligand-receptor knowledge base is a pre-integrated biological database containing known ligand-receptor pairs. A ligand is a signaling molecule that can be secreted by cells or expressed on the membrane, while a receptor is a protein that specifically recognizes and binds to a ligand, thereby triggering an intracellular response. The preset conditions are rules used to screen ligand-receptor pairs: the ligand gene must appear in the node set of the source cell consensus network, and the receptor gene must appear in the node set of the target cell consensus network. The candidate ligand-receptor pair list is a list where each item is a ligand-receptor pair that meets the preset conditions, containing all possible communication events between the two cell types currently being analyzed. The terminal queries the ligand-receptor knowledge base to obtain all known ligand-receptor pairs. Each ligand-receptor gene pair in the knowledge base is matched with all node genes in both the source and target cell consensus networks. Ligand-receptor pairs where the ligand gene can be found in the source cell network and the receptor gene can be found in the target cell network are retained. All successfully matched ligand-receptor pairs are collected to form the candidate ligand-receptor pair list.

[0111] Step 502: For each receptor in the ligand-receptor pair list, search for a path to the node of the transcription factor in the target cell consensus network to obtain a receptor-transcription factor path list.

[0112] Specifically, the receptor-transcription factor pathway list is a list where each item represents a complete receptor-transcription factor pathway. For each receptor in the candidate list, zero, one, or more pathways leading to different transcription factors may be found. Here, a receptor-transcription factor pathway specifically refers to a continuous connection relationship in the target cell consensus network, starting from a receptor gene node, passing through a series of intermediate signaling molecules, and reaching a transcription factor gene node, describing the signal transduction process within the cell. For example, the terminal traverses each receptor gene in the candidate ligand-receptor pair list, using each receptor gene as a starting point, and performs a path search in the target cell consensus network graph using a breadth-first search, with a maximum path length limit of no more than 5 steps to ensure the biological rationality and interpretability of the path. During the search, all paths starting from that receptor gene and ending at any transcription factor node are recorded. All such paths found for all receptors are summarized to form the receptor-transcription factor pathway list.

[0113] All such pathways found for all receptors are compiled into a list of receptor-transcription factor pathways.

[0114] Step 503: In the target cell consensus network, query and extract all target genes regulated by transcription factors to obtain transcription factor-target gene pathways; and add the transcription factor-target gene pathways to the receptor-transcription factor pathway list to obtain an expanded signaling pathway list.

[0115] Specifically, the transcription factor-target gene pathway refers to the edge in the target cell consensus network where a transcription factor node directly points to the target gene node it regulates. This represents the final effect of the signaling pathway, i.e., the regulatory effect of the transcription factor on a specific gene. The extended signaling pathway list is formed by adding all directly regulated target genes to the endpoint of each pathway, based on the receptor-transcription factor pathway list. The terminal queries and extracts all existing transcription factor-target gene regulatory relationships in the target cell consensus network, traverses each pathway in the receptor-transcription factor pathway list, finds the transcription factor for each pathway, queries all target genes regulated by that transcription factor, and adds each transcription factor-target gene regulatory relationship as a new branch to the end of the original pathway. This expands a receptor-transcription factor pathway into multiple pathways from the receptor to different target genes. All expanded pathways are collected to form the extended signaling pathway list.

[0116] Step 504: Connect the candidate ligand-receptor pair list and the signaling pathway list to obtain a transcellular signaling pathway list.

[0117] The connection here refers to the logical splicing and assembly, linking upstream ligand-receptor pairs with downstream extended signaling pathways in biological order. The transcellular signaling pathway list is a list where each item is a complete, hypothetical transcellular signaling pathway; for example, the format could be [ligand], [receptor], {signal transduction pathway}, [transcription factor], {target gene}. The terminal iterates through each ligand-receptor pair in the candidate ligand-receptor pair list, searches the extended signaling pathway list for all pathways starting with that receptor, takes the ligand-receptor pair as the beginning of the pathway, and treats each corresponding extended pathway as the subsequent downstream part, logically concatenating them to summarize all assembled complete pathways, forming the transcellular signaling pathway list.

[0118] This embodiment reveals, through the complete construction of the signaling pathway from start to finish, which factors may lead to changes in gene expression due to transcellular signaling. This makes the biological significance of the pathway clearer and more complete, provides specific and verifiable hypotheses, and explains how two cell types may communicate through specific molecules and how this communication affects the gene expression program of the target cell.

[0119] In one embodiment, for each receptor in the ligand-receptor pair list, a path to the transcription factor node is searched in the target cell consensus network to obtain a receptor-transcription factor path list, including:

[0120] Step 601: Extract all receptor genes located on the target cell side from the candidate ligand-receptor pair list to generate a receptor gene list.

[0121] The candidate ligand-receptor pair list is a list in which each item contains a ligand gene and a receptor gene, representing a possible intercellular communication event. The receptor gene on the target cell side is specifically designated within each ligand-receptor pair, representing the receptor gene present in the target cell's consensus network and serving as the entry point for signal transmission from the extracellular space to the intracellular space. The receptor gene list is a deduplicated list of genes containing all unique receptor genes extracted from the candidate ligand-receptor pair list. This clarifies all starting points for subsequent downstream path searches within the target cell. The terminal iterates through each item in the candidate ligand-receptor pair list, extracting the receptor gene representing the target cell side. All extracted receptor genes are collected, and duplicate genes are removed to generate the receptor gene list.

[0122] Step 602: Using each receptor gene in the receptor gene list as a starting node, traverse the target cell consensus network, record the direct neighboring nodes of the initial node, and obtain the potential path set.

[0123] Specifically, the starting node is the gene node at the beginning of this path search, which is a receptor gene in the receptor gene list. Directly adjacent nodes are all nodes in the target cell consensus network that are directly connected to the starting node via an edge, representing proteins or molecules that directly interact with the receptor. The potential path set is a set whose members are paths of length 1 from a starting node to each of its directly adjacent nodes. The terminal sequentially selects a receptor gene from the receptor gene list as the starting node for this search. In the target cell consensus network, it locates this starting node and queries all nodes with which it has a directly interacting edge. For each directly adjacent node found, a path from the starting node to that adjacent node is recorded, forming the potential path set for that receptor. This process is repeated for each receptor gene in the list to obtain the potential path set for all receptor genes.

[0124] Step 603: Filter out the paths in the potential path set whose last node is a transcription factor to obtain the receptor-transcription factor path list.

[0125] Specifically, the last node in a path refers to the endpoint of that path. Transcription factors are proteins that bind to specific DNA sequences, thereby initiating or inhibiting gene transcription. They are key effector molecules in signaling pathways, translating signal transduction into changes in gene expression. The receptor-transcription factor path list is a subset selected from all potential path sets, where the last node of each path is a transcription factor, describing the process of signal transmission directly from the receptor to the transcription factor. The terminal iterates through every path in the set of all potential paths generated for all receptors. For each path, it checks against a known list of transcription factor genes to see if the last node is a transcription factor. If the node is a transcription factor, the path is retained. All paths that meet the criteria are then compiled to form the receptor-transcription factor path list.

[0126] This embodiment screened out a particularly important and efficient direct pathway in the signaling pathway. The signal is transmitted directly from the membrane receptor to activate or inhibit the transcription factor in the nucleus in one step, making the final constructed pathway more concise and with clear biological significance. Direct pathways are usually directly associated with rapid gene expression regulation, which enhances the biological significance of the signaling pathway.

[0127] In one embodiment, the global signal reference network is constructed using the following method:

[0128] Step 701: Obtain raw data on cell signaling interactions; and extract the participants and corresponding relationship types from the raw data to obtain a set of interaction data; the raw data includes at least one of the following: transcription factor-target gene regulatory relationships, protein interactions, metabolic responses, signal transduction events, and ligand-receptor pairing information.

[0129] Specifically, raw interaction data for cell signaling consists of unprocessed, raw records obtained from various public biological databases, describing various types of interactions between molecules. The participants in an interaction relationship refer to two or more biomolecular entities that interact in a single raw interaction record, uniquely identified by standard gene symbols or protein codes. The relationship type describes the specific manner of interaction between the participants. The interaction dataset is a structured list or dataset where each record contains a pair of participants and their specific relationship type. The terminal acquires various types of raw interaction data from multiple data sources, extracts information from each raw data record, identifies and extracts the interacting parties and the nature of the interaction, formats the extracted information into a standard record, and aggregates all records to form the initial interaction dataset.

[0130] Step 702: Standardize the action data set to obtain the standard action set.

[0131] The standard action set is a cleaned and standardized set of action data, in which gene and / or protein identifiers have been standardized and relation types have been merged into a standard classification. The terminal maps the various codes of participants to standard gene symbols, ensuring that the same molecule has the same identifier in all records. Relation types from different databases that have different expressions but the same meaning are merged into a unified standard type. Records that cannot be mapped or have ambiguous meanings are removed, resulting in a standard action set with internal consistency and no redundant identifiers.

[0132] Step 703: Based on the preset reliability criteria, remove records from the standard action set that do not meet the reliability requirements to obtain the reliable action set.

[0133] The preset reliability criteria are rules used to evaluate the credibility of an interaction record. Common criteria include experimental evidence (whether it is supported by reliable experimental methods), database support (whether it is included in multiple independent databases), and literature citation count (whether it is reported in multiple publications). The reliable interaction set is a subset of the standard interaction set, containing only interaction records that meet all preset reliability criteria, and thus has higher quality and credibility. The terminal evaluates the reliability of each record in the standard interaction set, checking its evidence type and the number of source databases. Each record is compared with the preset reliability criteria, and all records that do not meet the criteria are eliminated. The remaining high-confidence records are then combined into the reliable interaction set.

[0134] Step 704: Based on the reliable set of interactions, treat each gene or protein node and each interaction as an edge between nodes, and assign labels to the edges according to the type of interaction to obtain a preliminary signal reference network.

[0135] Specifically, nodes are the basic units in the graph. In this network, each node represents a biomolecular entity, specifically a unique gene or protein appearing in all records of the reliable interaction set. Edges are lines connecting two nodes, representing interactions; each edge corresponds to a record in the reliable interaction set. Edge labels are attributes assigned to edges, describing the type of interaction, directly derived from the relation types in the reliable interaction set. The preliminary signal reference network is a graph data structure where nodes are genes and / or proteins, edges are interactions between nodes, and edges are labeled with descriptions of interaction types. It is a unified biosignal network containing all information. The terminal extracts all unique molecules appearing in all records of the reliable interaction set and adds them as nodes to the network. It traverses each record in the reliable interaction set, connecting the nodes corresponding to the two participants in the record with an edge, and assigning the relation type in the record as the label to the newly created edge, thus generating the preliminary signal reference network.

[0136] Step 705: The preliminary signal reference network is divided into intracellular signal modules and intercellular signal modules to obtain a global signal reference network; the intracellular signal modules are used to characterize intracellular signal cascades; the intercellular signal modules are used to characterize ligand-receptor interaction pairs.

[0137] Specifically, the intracellular signaling module is a subnetwork of the preliminary signaling reference network, encompassing all interactions occurring within the cell. Edge types include transcription factor-target gene regulation, protein-protein interactions, signal transduction, and metabolic responses, describing intracellular signal transduction and regulatory pathways. The intercellular signaling module is another subnetwork of the preliminary signaling reference network, specifically containing all ligand-receptor pairing interactions, describing the communication interfaces between different cells. The global signaling reference network is a modularly partitioned signaling network composed of the intracellular and intercellular signaling modules, used for all subsequent analyses. The terminal classifies all edges in the preliminary signaling reference network. Based on the edge labels, it extracts all edges belonging to the ligand-receptor type and their connected nodes, forming the intercellular signaling module separately. All other remaining edges and their connected nodes form the intracellular signaling module. These two modules together constitute the global signaling reference network.

[0138] This embodiment generates a knowledge base that includes different signal modules, enabling subsequent analysis to efficiently and specifically call different parts of the knowledge base, thereby improving the processing speed of pathway reconstruction and solving the problem of the lack of a reliable and comprehensive prior knowledge base for analysis.

[0139] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0140] Based on the same inventive concept, this application also provides a single-cell signaling pathway reconstruction system for implementing the single-cell signaling pathway reconstruction method described above. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the single-cell signaling pathway reconstruction system provided below can be found in the limitations of the single-cell signaling pathway reconstruction method described above, and will not be repeated here.

[0141] In one exemplary embodiment, such as Figure 2 As shown, a single-cell signaling pathway reconstruction system 800 is provided, comprising:

[0142] The generation module 801 is used to extract sub-networks from the global signal reference network based on the intersection of highly expressed genes and hypervariable genes in each cell, and generate cell-specific signal networks; the nodes of the cell-specific signal networks are hypervariable genes, and the edges are the interactions between hypervariable genes.

[0143] Consensus module 802 is used to input cell-specific signaling networks into the graph Transformer model and generate consensus signaling networks;

[0144] The definition module 803 is used to define the consensus signal network of the corresponding source cell as the source cell consensus network and the consensus signal network of the corresponding target cell as the target cell consensus network.

[0145] Path module 804 is used to search for pathways between ligands in the source cell consensus network and receptors in the target cell consensus network based on the ligand-receptor knowledge base, and generate a list of cross-cell signaling pathways.

[0146] Furthermore, consensus module 802 is also used for:

[0147] Each node in the cell-specific signaling network is downsampled into a low-dimensional value to obtain the encoded node embedding vector; and the encoded node embedding vectors are aggregated into a cell embedding vector.

[0148] Based on cell type annotation information, cell embedding vectors belonging to the same cell type are aggregated to obtain consensus cell embedding vectors.

[0149] Based on the consensus cell embedding vector, a low-dimensional vector is generated for each node to obtain the inference node embedding vector; the inference node embedding vector represents the functional context of the node in the typical state of cell type.

[0150] Based on the embedding vectors of the inference nodes, the connection probabilities between each node are calculated using the following formula to obtain the probabilistic adjacency matrix:

[0151]

[0152] in, Let be the probability that there is an edge between node i and node j. Let i be the node embedding vector. Let j be the node embedding vector;

[0153] The consensus signal network is obtained by eliminating the interactions between nodes whose connection probability is less than a preset threshold.

[0154] Furthermore, the system also includes a modeling module for:

[0155] The training set is input into the encoder of the graph Transformer model to generate encoding node embedding vectors and cell embedding vectors; the training set includes cell-specific signaling networks and corresponding cell type labels.

[0156] The encoded nodes are embedded into vectors and input into the decoder. The connection probability between each node is calculated, and a reconstruction probability matrix is ​​generated.

[0157] The cell embedding vector is input into the classifier to obtain the cell type probability distribution;

[0158] Based on the reconstruction probability matrix, cell type probability distribution, adjacency matrix of cell-specific signal network, and cell type label, multi-task loss calculation is performed to obtain the joint loss value;

[0159] Based on the joint loss value, the parameters of each module of the graph Transformer model are updated to obtain the updated graph Transformer model.

[0160] Furthermore, the modeling module is also used for:

[0161] Based on the reconstruction probability matrix and the adjacency matrix of the cell-specific signaling network, the reconstruction loss is calculated using the following formula:

[0162]

[0163]

[0164] in, For positive sample reconstruction loss, It is the set of all pairs of nodes in the adjacency matrix that have an edge. For negative sample reconstruction loss, It is the set of all pairs of nodes in the adjacency matrix that do not have an edge.

[0165] Based on the cell type probability distribution and cell type labels, the classification loss is calculated using the following formula:

[0166]

[0167] in, The classification loss is represented by C, where C is the total number of cell types. For actual cell type labels, Let c be the probability that the cell belongs to type c.

[0168] The joint loss value is obtained by weighted summation of reconstruction loss and classification loss.

[0169] Furthermore, path module 804 is also used for:

[0170] The ligand-receptor knowledge base is queried to match all ligand-receptor pairs that meet the preset conditions, resulting in a list of candidate ligand-receptor pairs. The preset conditions are that the ligand gene exists in a node of the source cell consensus network and the receptor gene exists in a node of the target cell consensus network.

[0171] For each receptor in the ligand-receptor pair list, a path leading to the transcription factor node is searched in the target cell consensus network to obtain a receptor-transcription factor path list.

[0172] In the target cell consensus network, all target genes regulated by transcription factors are queried and extracted to obtain the transcription factor-target gene pathway; and the transcription factor-target gene pathway is added to the receptor-transcription factor pathway list to obtain an expanded signaling pathway list.

[0173] By concatenating the candidate ligand-receptor pair list and the signaling pathway list, a list of transcellular signaling pathways is obtained.

[0174] Furthermore, path module 804 is also used for:

[0175] Extract all receptor genes located on the target cell side from the candidate ligand-receptor pair list to generate a receptor gene list;

[0176] Starting with each receptor gene in the receptor gene list as a starting node, the target cell consensus network is traversed, and the direct neighboring nodes of the initial node are recorded to obtain a set of potential paths.

[0177] By filtering out the potential pathways whose last node is a transcription factor, a list of receptor-transcription factor pathways is obtained.

[0178] Furthermore, the generation module 801 is also used for:

[0179] Obtain raw data on cell signaling interactions; and extract the participants and corresponding relationship types from the raw data to obtain a set of interaction data; the raw data includes at least one of the following: transcription factor-target gene regulatory relationships, protein-protein interactions, metabolic responses, signal transduction events, and ligand-receptor pairing information;

[0180] The action data set is standardized to obtain the standard action set;

[0181] Based on the preset reliability criteria, records that do not meet the reliability requirements in the standard action set are removed to obtain the reliable action set;

[0182] Based on a reliable set of interactions, each gene or protein node and each interaction are treated as an edge between nodes, and the edges are labeled according to the type of interaction to obtain a preliminary signal reference network.

[0183] The initial signal reference network is divided into intracellular signal modules and intercellular signal modules to obtain a global signal reference network; the intracellular signal modules are used to characterize intracellular signal cascades; and the intercellular signal modules are used to characterize ligand-receptor interaction pairs.

[0184] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the single-cell signaling pathway reconstruction method as described above.

[0185] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0186] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0187] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for reconstructing a single-cell signaling pathway, characterized in that, The method includes: Based on the intersection of highly expressed genes and hypervariable genes in each cell, a sub-network is extracted from the global signal reference network to generate a cell-specific signal network; the nodes of the cell-specific signal network are the hypervariable genes, and the edges are the interactions between the hypervariable genes. The cell-specific signaling network is input into a graph Transformer model to generate a consensus signaling network. The consensus signal network corresponding to the source cell is defined as the source cell consensus network, and the consensus signal network corresponding to the target cell is defined as the target cell consensus network; Based on the ligand-receptor knowledge base, the pathways between ligands in the source cell consensus network and receptors in the target cell consensus network are searched to generate a list of transcellular signaling pathways.

2. The method according to claim 1, characterized in that, The graph Transformer model generates the consensus signaling network using the following method: Each node in the cell-specific signaling network is downsampled into a low-dimensional value to obtain an encoded node embedding vector; and the encoded node embedding vectors are aggregated into a cell embedding vector. Based on cell type annotation information, the cell embedding vectors belonging to the same cell type are aggregated to obtain consensus cell embedding vectors; Based on the consensus cell embedding vector, a low-dimensional vector is generated for each node to obtain the inference node embedding vector; The inference node embedding vector represents the functional context of the node in a typical state of cell type; Based on the inference node embedding vectors, the connection probabilities between the nodes are calculated using the following formula to obtain the probabilistic adjacency matrix: in, Let be the probability that there is an edge between node i and node j. Let i be the node embedding vector. Let j be the node embedding vector; The consensus signal network is obtained by eliminating the interactions between nodes whose connection probability is less than a preset threshold.

3. The method according to claim 2, characterized in that, The graph Transformer model is trained using the following method: The training set is input into the encoder of the graph Transformer model to generate the encoding node embedding vector and the cell embedding vector; the training set includes the cell-specific signaling network and the corresponding cell type label. The encoded nodes are embedded into vectors and input into the decoder to calculate the connection probabilities between each node, thereby generating a reconstruction probability matrix. The cell embedding vector is input into a classifier to obtain the cell type probability distribution; Based on the reconstruction probability matrix, the cell type probability distribution, the adjacency matrix of the cell-specific signal network, and the cell type label, multi-task loss calculation is performed to obtain the joint loss value; Based on the joint loss value, the parameters of each module of the graph Transformer model are updated to obtain the updated graph Transformer model.

4. The method according to claim 3, characterized in that, The process of calculating a joint loss value based on the reconstructed probability matrix, the cell type probability distribution, the adjacency matrix of the cell-specific signal network, and the cell type label includes: Based on the reconstruction probability matrix and the adjacency matrix of the cell-specific signaling network, the reconstruction loss is calculated using the following formula: in, For positive sample reconstruction loss, It is the set of all pairs of nodes in the adjacency matrix that have an edge. For negative sample reconstruction loss, It is the set of all pairs of nodes in the adjacency matrix that do not have an edge. Based on the cell type probability distribution and the cell type label, the classification loss is calculated using the following formula: in, The classification loss is represented by C, where C is the total number of cell types. For actual cell type labels, Let c be the probability that the cell belongs to type c. The joint loss value is obtained by weighted summing of the reconstruction loss and the classification loss.

5. The method according to claim 1, characterized in that, The method, based on a ligand-receptor knowledge base, searches for pathways between ligands in the source cell consensus network and receptors in the target cell consensus network to generate a list of cross-cell signaling pathways, including: The ligand-receptor knowledge base is queried, and all ligand-receptor pairs that meet the preset conditions are matched to obtain a list of candidate ligand-receptor pairs; the preset conditions are that the ligand gene exists in the node of the source cell consensus network and the receptor gene exists in the node of the target cell consensus network. For each receptor in the ligand-receptor pair list, a path to the transcription factor node is searched in the target cell consensus network to obtain a receptor-transcription factor path list. In the target cell consensus network, all target genes regulated by the transcription factors are queried and extracted to obtain the transcription factor-target gene pathway; and the transcription factor-target gene pathway is added to the receptor-transcription factor pathway list to obtain an expanded signal pathway list. By concatenating the candidate ligand-receptor pair list and the signaling pathway list, the transcellular signaling pathway list is obtained.

6. The method according to claim 5, characterized in that, For each receptor in the ligand-receptor pair list, a path to the transcription factor node is searched in the target cell consensus network to obtain a receptor-transcription factor path list, including: Extract all receptor genes located on the target cell side from the candidate ligand-receptor pair list to generate a receptor gene list; By taking each receptor gene in the receptor gene list as a starting node, the target cell consensus network is traversed, and the direct neighboring nodes of the initial node are recorded to obtain a set of potential paths. The path whose last node in the potential path set is the transcription factor is selected to obtain the receptor-transcription factor path list.

7. The method according to claim 1, characterized in that, The global signal reference network is constructed using the following method: Obtain raw data on cell signaling interactions; and extract the participants and corresponding relationship types from the raw data to obtain a set of interaction data; the raw data includes at least one of transcription factor-target gene regulatory relationships, protein interactions, metabolic responses, signal transduction events, and ligand-receptor pairing information; The action data set is standardized to obtain a standard action set; Based on the preset reliability criteria, records that do not meet the reliability requirements in the standard action set are removed to obtain the reliable action set; Based on the reliable set of interactions, each node of a gene or protein is treated as an edge between the nodes, and the edges are labeled according to the type of the interaction to obtain a preliminary signal reference network. The preliminary signal reference network is divided into intracellular signal modules and intercellular signal modules to obtain the global signal reference network; the intracellular signal modules are used to characterize intracellular signal cascades; and the intercellular signal modules are used to characterize ligand-receptor interaction pairs.

8. A single-cell signaling pathway reconstruction system, characterized in that, The system includes: The generation module is used to extract sub-networks from the global signal reference network based on the intersection of highly expressed genes and hypervariable genes in each cell, and generate a cell-specific signal network; the nodes of the cell-specific signal network are the hypervariable genes, and the edges are the interactions between the hypervariable genes. The consensus module is used to input the cell-specific signaling network into the graph Transformer model to generate a consensus signaling network; The definition module is used to define the consensus signal network corresponding to the source cell as the source cell consensus network and the consensus signal network corresponding to the target cell as the target cell consensus network. The path module is used to search for pathways between ligands in the source cell consensus network and receptors in the target cell consensus network based on the ligand-receptor knowledge base, and generate a list of cross-cell signaling pathways.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images